Nathaniel Smith, a Research Associate at the University of Edinburgh School of Informatics, will give a talk at 12:00 on Wednesday, Feb. 26, in Tobin 521B. The title of his talk is ''Building a Bayesian bridge between the physics and the phenomenology of social interaction.’’ An abstract follows.
What is word meaning, and where does it live? Both naive intuition andscientific theories in fields such as discourse analysis and socio-and cognitive linguistics place word meanings, at least in part,outside the head: in important ways, they are properties of speechcommunities rather than individual speakers. Yet, from aneuroscientific perspective, we know that actual speakers andlisteners have no access to such consensus meanings: the physicalprocesses which generate word tokens in usage can only depend directlyon the idiosyncratic goals, history, and mental state of a singleindividual. It is not clear how these perspectives can be reconciled.This gulf is thrown into sharp perspective by current Bayesian modelsof language processing: models of learning have taken the formerperspective, and models of pragmatic inference and implicature havetaken the latter. As a result, these two families of models, thoughbuilt using the same mathematical framework and often by the samepeople, turn out to contain formally incompatible assumptions.
Here, I'll present the first Bayesian model which can simultaneouslylearn word meanings and perform pragmatic inference. In addition tocapturing standard phenomena in both of these literatures, it givesinsight into how the literal meaning of words like "some" can beacquired from observations of pragmatically strengthened uses, andprovides a theory of how novel, task-appropriate linguisticconventions arise and persist within a single dialogue, such as occursin the well-known phenomenon of lexical alignment. Over longer timescales such effects should accumulate to produce language change;however, unlike traditional iterated learning models, our simulatedagents do not converge on a sample from their prior, but instead showan emergent bias towards belief in more useful lexicons. Our modelalso makes the interesting prediction that different classes ofimplicature should be differentially likely to conventionalize overtime. Finally, I'll argue that the mathematical "trick" needed toconvince word learning and pragmatics to work together in the samemodel is in fact capturing a real truth about the psychologicalmechanisms needed to support human culture, and, more speculatively,suggest that it may point the way towards a general mechanism forreconciling qualitative, externalist theories of social interactionwith quantitative, internalist models of low-level perception andaction, while preserving the key claims of both approaches.